/projects/

Deep Learning for Accessibility: Detection and Segmentation of Regions of Interest for Sign Language Recognition Systems

Supervisors: Dr. Frederico Gadelha and MSc. Tamires Rezende

First thesis presented to the Undergraduate Program in Systems Engineering of the Federal University of Minas Gerais in partial fulfillment of the requirements for the degree of Bachelor of Science in Systems Engineering.

The focus of this study is to understand how Deep Learning techniques can be applied to enhance the performance of sign language recognition systems.

Thesis I: [PDF] [Presentation]

Thesis II: [PDF] [Presentation]

Giulia Zanon, Rúbia Guerra, et al. (2019)

Development of a Database of Libras’ Signs for Machine Learning: 3DCNN Case Study

A recurrent problem in Brazilian Sign Language (Libras) recognition is the absence of a robust dataset that allows the validation of different methodologies. This work presents a new dataset for Libras and its respective recording procedure. The first available version contains 20 signs, recorded 5 times by 10 different signers, making up 1000 recordings. A study case with the new data utilizing a 3D Convolutional Neural Network for sign recognition is also presented, employing summarization and data augmentation techniques. The network implemented achieved an average accuracy of 72,6%.

[PDF] PT-BR

Rúbia Guerra, Tamires Rezende, Frederico Gadelha, et al. (2018)

Facial Expression Analysis in Brazilian Sign Language for Sign Recognition

Sign language is one of the main forms of communication used by the deaf community. The language’s smallest unit, a “sign”, comprises a series of intricate manual and facial gestures. As opposed to speech recognition, sign language recognition (SLR) lags behind, presenting a multitude of open chal- lenges because this language is visual-motor. This paper aims to explore two novel approaches in feature extraction of facial expressions in SLR, and to pro- pose the use of Random Forest (RF) in Brazilian SLR as a scalable alternative to Support Vector Machines (SVM) and k-Nearest Neighbors (k-NN). Results show that RF’s performance is at least comparable to SVM’s and k-NN’s, and validate non-manual parameter recognition as a consistent step towards SLR.

[PDF]

Lucas Chaves, Mateus Rodrigues, Rúbia Guerra, et al. (2016)

Analysis of Quality of Life of Electrical Engineering Students

This paper analyzes factors of the academic life of students in the undergraduate program in Electrical Engineering at the Federal University of Minas Gerais and their influence on students' quality of life and performance. The data examined were obtained through a questionnaire composed of three research instruments, aiming to characterize the student body with regards to sociodemographic aspects, quality of life and overall satisfaction with the program. Statistical analyses of the responses (n=184 enrolled students) made it possible to compare the different factors that affect the quality of life of students.

[PDF] PT-BR